Entropic Annealing in Particle Filtering

• Entropic Annealing is an adaptive method that adjusts resampling temperatures based on entropy measures to prevent premature exploitation in particle filters.
• It monitors the normalized effective sample size (ESS) to dynamically control variance in particle-based Monte Carlo searches, preserving solution diversity.
• Studies show that an ESS-based temperature schedule outperforms linear and entropy-based methods, leading to significant improvements on challenging inference benchmarks.

Entropic Annealing (EA) denotes a family of entropy-guided annealing procedures in which an entropy, relative entropy, effective sample size, or entropic regularization parameter controls the transition from exploratory to exploitative behavior. In recent inference-time scaling for LLMs, the term refers specifically to the entropy-triggered resampling intervention inside Entropic Particle Filtering (ePF), introduced to mitigate premature exploitation and particle impoverishment in particle-based Monte Carlo search for mathematical reasoning (Giannone et al., 7 Oct 2025). In other literatures, the same name, or a closely related usage, refers to adaptive temperature ladders based on relative entropy, finite-temperature maximum-entropy inference, entropy-production-minimizing simulated annealing, and entropic-regularization schedules in optimal transport.

1. Terminological scope

The cited literature uses the expression “Entropic Annealing” for several distinct but related constructions. All of them treat entropy-like quantities as control variables rather than as passive diagnostics (Habeck, 2015, Chancellor et al., 2015, Herty et al., 17 Apr 2025, Albert, 2015, Zhang et al., 15 Jun 2026, Delalande, 2021, Jain et al., 13 Apr 2026).

Domain	Entropy/KL object	EA meaning
Particle filtering for ITS	Shannon entropy or normalized ESS of resampling weights	Dynamic flattening of resampling softmax
Ensemble annealing	Relative entropy between successive ensembles	Adaptive temperature ladder
Programmable annealer inference	Shannon entropy of Boltzmann posterior	Finite-temperature maximum-entropy decoding
Kinetic simulated annealing	Relative entropy to an instantaneous Gibbs state	Closed-loop entropy-based cooling
Bayesian simulated annealing	Entropy production rate	Adaptive annealing of ABC ensembles
Entropic OT / CPD	Entropy regularization or variance-induced temperature	Automatic coarse-to-fine correspondence sharpening
Semi-discrete OT	Entropic regularization parameter $\epsilon$	$\epsilon$-scaling with warm starts
Sparse localization	Entropy factorization along an annealing path	Local-to-global transfer of mLSI

This multiplicity is important. In contemporary large-language-model inference-time scaling, EA most commonly denotes the specific entropy-based resampling mechanism of ePF, whereas in statistical physics and optimal transport it more often denotes adaptive temperature or regularization schedules.

2. EA in particle-based Monte Carlo for inference-time scaling

In “Mitigating Premature Exploitation in Particle-based Monte Carlo for Inference-Time Scaling,” Particle Filtering (PF) is described as a strong inference-time scaling method for complex mathematical reasoning tasks, but one that is vulnerable when guided by process reward models (PRMs) that assign overconfident scores early in the reasoning process. The resulting failure mode is premature exploitation: PF myopically commits to locally promising trajectories, prunes potentially correct hypotheses, and converges to suboptimal solutions. This is identified with particle impoverishment and is especially severe under constrained computational budgets (Giannone et al., 7 Oct 2025).

The same work isolates two root causes. The first is a lack of diversity in the particle set due to overconfident resampling. The second is the consequent inability to assess the potential of a reasoning path. ePF addresses these with two additions to PF: Entropic Annealing (EA) and Look-ahead Modulation (LaM). EA is the diversity-preserving component. It monitors search diversity via entropy and, when diversity drops, dynamically anneals the resampling distribution so that exploration is preserved until sufficient information is gathered. LaM is a distinct enhancement that evaluates a state’s potential based on its successors (Giannone et al., 7 Oct 2025).

Within this formulation, EA is not a generic temperature schedule over model logits or tokens. It is a targeted intervention on the resampling distribution of a particle filter. Its purpose is variance control in the particle weights, with the operational goal of preventing early collapse of the particle set.

3. Diversity diagnostics, activation criterion, and annealed resampling

EA tracks the normalized Shannon entropy of the resampling weights at step $t$,

$$H_n(t) \;=\; -\,\frac{1}{\log N}\sum_{i=1}^N w^i_t\,\log w^i_t \quad\in[0,1],$$

or equivalently the normalized effective sample size

$$ESS_n(t) \;=\; \frac{1}{N}\,\frac{1}{\sum_{i=1}^N (w^i_t)^2} \;=\; \frac{1}{N}\,ESS(t), \qquad ESS(t)=\frac1{\sum_i(w^i_t)^2}\in[1,N].$$

Low $H_n$ or $ESS_n$ indicates that very few particles carry the mass. In the paper’s implementation, EA is activated whenever diversity falls below a threshold $\tau$, using

$$ESS_n(t)\;\le\;\tau \qquad\text{with}\qquad \tau=0.5$$

for the first half of the trajectory. The summary notes that one could equivalently use $H_n(t)\le\tau$ (Giannone et al., 7 Oct 2025).

When $\epsilon$0, EA introduces a temperature parameter $\epsilon$1 and recomputes the normalized weights by “heating up” the PRM log-scores $\epsilon$2: $\epsilon$3 As $\epsilon$4 decreases, the softmax flattens, reducing variance and preserving more particles. The ESS-based schedule reported as most effective is

$\epsilon$5

Early in the trajectory and under low diversity, this yields strong flattening; as $\epsilon$6, it forces $\epsilon$7, recovering the original PF softmax for exploitation. The same work compares this schedule to linear decay,

$\epsilon$8

and to an entropy-based rule,

$\epsilon$9

and reports that the ESS-based rule is most effective (Giannone et al., 7 Oct 2025).

Integrated into PF, the workflow is: propagate and score particles, compute vanilla softmax weights from PRM log-scores, evaluate $t$0, anneal the weights if the threshold is crossed, resample via systematic or stratified resampling, and reset weights to $t$1. The key hyperparameters are the diversity threshold $t$2 (default $t$3), the schedule choice, the resampling method, and the early-stop fraction that applies EA for the first $t$4 of steps. The summary further states that systematic or stratified resampling has lower variance than multinomial, and recommends using $t$5 so that the procedure reverts at worst to standard PF (Giannone et al., 7 Oct 2025).

4. Variance control, theoretical properties, and reported performance

The theoretical motivation for EA is given in Monte Carlo variance terms. High weight variance degrades the posterior approximation, and degeneracy leads to impoverishment. By contrast, high entropy in the resampling distribution corresponds to low variance and a high effective sample size. The summary states the relation

$t$6

On this view, flattening the weights when $t$7 grows is an online variance-reduction mechanism. The same source further states that, in the limit of infinite particles, EA recovers standard PF, while with finite $t$8 it provably reduces worst-case weight variance at each resampling step (Giannone et al., 7 Oct 2025).

The paper’s ablations report that the ESS-based temperature rule outperforms the linear and entropy-based schedules, and that effective-sample-size plots confirm that EA keeps ESS high in early steps. The summary explicitly links these high-ESS early phases to higher pass@1. Reported benchmark comparisons isolate EA alone by comparing PF with ePF under “EA only, LaM off,” using Qwen2.5-1.5B-Instruct, $t$9, and uniform weighting (Giannone et al., 7 Oct 2025).

Benchmark	PF	ePF (EA only)
MATH500	60.15%	66.42% (+6.3)
DEEPMATH	22.65%	25.00% (+2.35)
OMNIMATH	8.59%	10.15% (+1.56)
AIME 2024	9.00%	11.20% (+2.2)
AIME 2025	6.66%	10.82% (+4.16)

At the full ePF level, where EA is combined with LaM, the system is reported to achieve up to a $$H_n(t) \;=\; -\,\frac{1}{\log N}\sum_{i=1}^N w^i_t\,\log w^i_t \quad\in[0,1],$$0 relative improvement in task reward on several challenging math benchmarks. This places EA within a broader strategy that balances exploration of diverse solution spaces with exploitation of high-reward regions, but the quantitative table above isolates the effect of EA itself (Giannone et al., 7 Oct 2025).

5. Thermodynamic, Bayesian, and optimization formulations

In statistical physics, Habeck’s “Ensemble annealing of complex physical systems” defines annealing through a sequence of inverse temperatures $$H_n(t) \;=\; -\,\frac{1}{\log N}\sum_{i=1}^N w^i_t\,\log w^i_t \quad\in[0,1],$$1 chosen so that the Kullback–Leibler divergence between successive ensembles is constant: $$H_n(t) \;=\; -\,\frac{1}{\log N}\sum_{i=1}^N w^i_t\,\log w^i_t \quad\in[0,1],$$2 The method simultaneously simulates the system and estimates its density of states using histogram reweighting (“WHAM”). In the 2D Ising model with $$H_n(t) \;=\; -\,\frac{1}{\log N}\sum_{i=1}^N w^i_t\,\log w^i_t \quad\in[0,1],$$3, $$H_n(t) \;=\; -\,\frac{1}{\log N}\sum_{i=1}^N w^i_t\,\log w^i_t \quad\in[0,1],$$4 walkers, and $$H_n(t) \;=\; -\,\frac{1}{\log N}\sum_{i=1}^N w^i_t\,\log w^i_t \quad\in[0,1],$$5, the procedure reaches $$H_n(t) \;=\; -\,\frac{1}{\log N}\sum_{i=1}^N w^i_t\,\log w^i_t \quad\in[0,1],$$6 after $$H_n(t) \;=\; -\,\frac{1}{\log N}\sum_{i=1}^N w^i_t\,\log w^i_t \quad\in[0,1],$$7 steps, and the relative error in the microcanonical entropy $$H_n(t) \;=\; -\,\frac{1}{\log N}\sum_{i=1}^N w^i_t\,\log w^i_t \quad\in[0,1],$$8 is reported as better than $$H_n(t) \;=\; -\,\frac{1}{\log N}\sum_{i=1}^N w^i_t\,\log w^i_t \quad\in[0,1],$$9 over most of $$ESS_n(t) \;=\; \frac{1}{N}\,\frac{1}{\sum_{i=1}^N (w^i_t)^2} \;=\; \frac{1}{N}\,ESS(t), \qquad ESS(t)=\frac1{\sum_i(w^i_t)^2}\in[1,N].$$0. The same framework is also illustrated on a 10-state Potts model and a protein Gō-model, where the annealing schedule slows near the critical region and the estimated density of states agrees with a long parallel-tempering reference (Habeck, 2015).

In hardware-based inference, “Maximum-Entropy Inference with a Programmable Annealer” uses “Entropic Annealing” to mean finite-temperature or maximum-entropy inference. Instead of selecting only the ground state

$$ESS_n(t) \;=\; \frac{1}{N}\,\frac{1}{\sum_{i=1}^N (w^i_t)^2} \;=\; \frac{1}{N}\,ESS(t), \qquad ESS(t)=\frac1{\sum_i(w^i_t)^2}\in[1,N].$$1

the method samples from a Boltzmann distribution

$$ESS_n(t) \;=\; \frac{1}{N}\,\frac{1}{\sum_{i=1}^N (w^i_t)^2} \;=\; \frac{1}{N}\,ESS(t), \qquad ESS(t)=\frac1{\sum_i(w^i_t)^2}\in[1,N].$$2

at a nonzero decoding temperature and decodes using thermal averages. For the programmable Josephson-junction array studied there, the D-Wave chip is operated at $$ESS_n(t) \;=\; \frac{1}{N}\,\frac{1}{\sum_{i=1}^N (w^i_t)^2} \;=\; \frac{1}{N}\,ESS(t), \qquad ESS(t)=\frac1{\sum_i(w^i_t)^2}\in[1,N].$$3 with anneal time $$ESS_n(t) \;=\; \frac{1}{N}\,\frac{1}{\sum_{i=1}^N (w^i_t)^2} \;=\; \frac{1}{N}\,ESS(t), \qquad ESS(t)=\frac1{\sum_i(w^i_t)^2}\in[1,N].$$4, and the decoded spin orientation is taken as the sign of the empirical thermal average. The experiments report that maximum-entropy decoding can in certain cases give competitive and even slightly better bit-error-rates than zero-temperature maximum-likelihood decoding. The same work develops a microscopic spin-sign-transition analysis and argues that the limiting factor is likely to be control errors of $$ESS_n(t) \;=\; \frac{1}{N}\,\frac{1}{\sum_{i=1}^N (w^i_t)^2} \;=\; \frac{1}{N}\,ESS(t), \qquad ESS(t)=\frac1{\sum_i(w^i_t)^2}\in[1,N].$$5–$$ESS_n(t) \;=\; \frac{1}{N}\,\frac{1}{\sum_{i=1}^N (w^i_t)^2} \;=\; \frac{1}{N}\,ESS(t), \qquad ESS(t)=\frac1{\sum_i(w^i_t)^2}\in[1,N].$$6 of the coupling scale rather than failure to reach equilibrium (Chancellor et al., 2015).

In kinetic simulated annealing, Herty and Zanella formulate an entropy-based cooling law in an extended phase space with densities $$ESS_n(t) \;=\; \frac{1}{N}\,\frac{1}{\sum_{i=1}^N (w^i_t)^2} \;=\; \frac{1}{N}\,ESS(t), \qquad ESS(t)=\frac1{\sum_i(w^i_t)^2}\in[1,N].$$7 over states and $$ESS_n(t) \;=\; \frac{1}{N}\,\frac{1}{\sum_{i=1}^N (w^i_t)^2} \;=\; \frac{1}{N}\,ESS(t), \qquad ESS(t)=\frac1{\sum_i(w^i_t)^2}\in[1,N].$$8 over temperatures. They measure the mismatch of $$ESS_n(t) \;=\; \frac{1}{N}\,\frac{1}{\sum_{i=1}^N (w^i_t)^2} \;=\; \frac{1}{N}\,ESS(t), \qquad ESS(t)=\frac1{\sum_i(w^i_t)^2}\in[1,N].$$9 to an instantaneous Gibbs state $H_n$0 by the relative entropy

$H_n$1

and choose a state-dependent feedback control

$H_n$2

so as to obtain

$H_n$3

Their numerical experiments use $H_n$4, $H_n$5, $H_n$6, $H_n$7, and $H_n$8, and report exponential entropy decay, faster temperature drop, and rapid concentration at the global minimizer relative to classical logarithmic cooling (Herty et al., 17 Apr 2025).

In Bayesian computation, Albert’s “A Simulated Annealing Approach to Bayesian Inference” develops a particle ensemble in parameter and output space, interprets the distance to observed data as an energy, and reduces the associated temperature so that entropy production is minimized. Under an endoreversibility assumption, the entropy-production rate is

$H_n$9

For a special metric choice one obtains $ESS_n$0 and an adaptive schedule with asymptotic behavior

$ESS_n$1

In the uninformative-prior limit with infinitely fast mixing and $ESS_n$2, the optimal fast schedule becomes

$ESS_n$3

The procedure is presented as a generic route to Bayesian posterior approximation without explicit likelihood-density evaluation (Albert, 2015).

6. Optimal transport, localization, and field-specific meanings

In optimal transport and point-cloud registration, entropic annealing often takes the form of decreasing an entropic regularization parameter or its equivalent. “Sinkhorn-CPD: Robust point cloud registration via unbalanced entropic optimal transport” replaces CPD’s target-side marginal constraint with dual Kullback–Leibler penalties and solves a fully unbalanced entropic OT problem by generalized Sinkhorn iterations. There, the CPD variance $ESS_n$4 plays exactly the role of the entropic regularization parameter: large $ESS_n$5 produces diffuse correspondences, and small $ESS_n$6 produces sharp correspondences. The automatic update

$ESS_n$7

therefore acts as a data-driven annealing schedule. The reported experiments state that this schedule yields sub-degree rotation error even at $ESS_n$8 outliers on the Stanford Bunny and achieves $ESS_n$9 registration recall on ModelNet40 with the criterion $\tau$0, $\tau$1, using a single setting $\tau$2 (Zhang et al., 15 Jun 2026).

In semi-discrete entropic optimal transport, entropic annealing is also called $\tau$3-scaling. Delalande studies the primal problem

$\tau$4

and justifies a schedule $\tau$5 with warm starts. The core result is “better-than-Lipschitz” stability of the Sinkhorn potentials: $\tau$6 This motivates geometric schedules

$\tau$7

which balance the number of stages against the warm-start error and reduce overall complexity relative to solving directly at the smallest $\tau$8 (Delalande, 2021).

A different, more structural use appears in “Entropic independence via sparse localization,” where annealing refers to a local-to-global path of measures

$\tau$9

interpolating between a hard target and an easy endpoint. Entropic independence is formulated through the entropy-factorization inequality

$$ESS_n(t)\;\le\;\tau \qquad\text{with}\qquad \tau=0.5$$0

and sparse localization shows that control of $$ESS_n(t)\;\le\;\tau \qquad\text{with}\qquad \tau=0.5$$1-independence only for sparse pinnings can still imply entropic independence. The main theorem quoted in the summary states that if a measure on the $$ESS_n(t)\;\le\;\tau \qquad\text{with}\qquad \tau=0.5$$2-slice has one-site marginals bounded below by $$ESS_n(t)\;\le\;\tau \qquad\text{with}\qquad \tau=0.5$$3 and is $$ESS_n(t)\;\le\;\tau \qquad\text{with}\qquad \tau=0.5$$4-$$ESS_n(t)\;\le\;\tau \qquad\text{with}\qquad \tau=0.5$$5-independent on $$ESS_n(t)\;\le\;\tau \qquad\text{with}\qquad \tau=0.5$$6, then it is $$ESS_n(t)\;\le\;\tau \qquad\text{with}\qquad \tau=0.5$$7-entropically independent with

$$ESS_n(t)\;\le\;\tau \qquad\text{with}\qquad \tau=0.5$$8

and the $$ESS_n(t)\;\le\;\tau \qquad\text{with}\qquad \tau=0.5$$9 loss is sharp in that framework (Jain et al., 13 Apr 2026).

Taken together, these usages show that “Entropic Annealing” is a field-dependent term rather than a single canonical algorithm. Depending on context, the annealed quantity may be a resampling softmax temperature $H_n(t)\le\tau$0, a physical inverse temperature $H_n(t)\le\tau$1, a decoding temperature $H_n(t)\le\tau$2, an entropic OT regularization $H_n(t)\le\tau$3, a CPD variance $H_n(t)\le\tau$4, or a schedule determined by entropy production. Likewise, the relevant entropy may be the Shannon entropy of particle weights, the KL divergence between successive ensembles, the entropy of a Boltzmann posterior, relative entropy to an instantaneous Gibbs state, or an entropy-factorization functional. What unifies these constructions is the use of entropy or KL structure to regulate annealing adaptively rather than through a fixed, purely exogenous schedule.